def infererence(args):
groups = 8
print('Loading image')
image = Image.open(args.image)
print('Preprocessing')
transformer = get_transformer()
input_data = preprocess(image, transformer)
print('input_data ', input_data.shape)
#conv layer
w = np.load('./data/' + 'module.conv1.weight.npy')
b = np.load('./data/' + 'module.conv1.bias.npy')
conv_layer = Convolution(w, b, stride=2, pad=1)
out = conv_layer.forward(input_data)
#savetxt('./dump/' + 'conv1_out.txt', out)
#max pooling
maxpool_layer = Pooling(3, 3, 2, 1)
out = maxpool_layer.forward(out)
#savetxt('./dump/' + 'maxpool_out.txt', out)
out = stage_shuffle(out, stage2_str, 3, groups)
#savetxt('./dump/' + 'stage2.txt', out)
out = stage_shuffle(out, stage3_str, 7, groups)
#savetxt('./dump/' + 'stage3.txt', out)
out = stage_shuffle(out, stage4_str, 3, groups)
#savetxt('./dump/' + 'stage4.txt', out)
h, w = out.shape[-2:]
avgpool_layer = AVGPooling(h, w, 1, 0)
out = avgpool_layer.forward(out).reshape(1, -1)
w = np.load('./data/' + 'module.fc.weight.npy')
b = np.load('./data/' + 'module.fc.bias.npy')
w = w.transpose(1, 0)
fc_layer = Affine(w, b)
out = fc_layer.forward(out)
softmax_layer = Softmax()
out = softmax_layer.forward(out).reshape(-1)
result = []
with open(args.idx_to_class) as json_file:
json_data = json.load(json_file)
'''
for key in json_data:
print(key, json_data[key])
'''
for i in range(0, 1000):
item = (out[i], json_data[str(i)])
result.append(item)
result = sorted(result, key=lambda item: item[0], reverse=True)
for i in range(0, 10):
print(result[i])
def stage_shuffle(input_data, stage, repeat_num, groups):
avgpool_layer = AVGPooling(3, 3, 2, 1)
residual = avgpool_layer.forward(input_data)
#savetxt('./dump/' + 'avg_pool.txt', residual)
w = np.load(stage + '0.g_conv_1x1_compress.conv1x1.weight.npy')
b = np.load(stage + '0.g_conv_1x1_compress.conv1x1.bias.npy')
if 'Stage2' in stage:
conv_layer = Convolution(w, b, stride=1, pad=0)
else:
conv_layer = GroupConvolution(w, b, stride=1, pad=0, groups=groups)
out = conv_layer.forward(input_data)
out_N, out_C, out_H, out_W = out.shape
gamma = np.load(stage +
'0.g_conv_1x1_compress.batch_norm.weight.npy').reshape(
(-1, 1))
beta = np.load(stage +
'0.g_conv_1x1_compress.batch_norm.bias.npy').reshape(
(-1, 1))
mean = np.load(
stage + '0.g_conv_1x1_compress.batch_norm.running_mean.npy').reshape(
(-1, 1))
var = np.load(stage +
'0.g_conv_1x1_compress.batch_norm.running_var.npy').reshape(
(-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out = bn_layer.forward(out.reshape(out_C, -1), train_flg=False)
relu_layer = Relu()
out = relu_layer.forward(out).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + '1x1_comp.txt', out)
out = channel_shuffle(out, groups)
#savetxt('./dump/' + 'channel_shuffle.txt', out)
w = np.load(stage + '0.depthwise_conv3x3.weight.npy').transpose(1, 0, 2, 3)
b = np.load(stage + '0.depthwise_conv3x3.bias.npy')
dwconv_layer = DWConvolution(w, b, stride=2, pad=1)
out = dwconv_layer.forward(out)
#savetxt('./dump/' + 'dwconv.txt', out)
gamma = np.load(stage + '0.bn_after_depthwise.weight.npy').reshape((-1, 1))
beta = np.load(stage + '0.bn_after_depthwise.bias.npy').reshape((-1, 1))
mean = np.load(stage + '0.bn_after_depthwise.running_mean.npy').reshape(
(-1, 1))
var = np.load(stage + '0.bn_after_depthwise.running_var.npy').reshape(
(-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out_N, out_C, out_H, out_W = out.shape
out = bn_layer.forward(out.reshape(out_C, -1),
train_flg=False).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + 'after_bn.txt', out)
w = np.load(stage + '0.g_conv_1x1_expand.conv1x1.weight.npy')
b = np.load(stage + '0.g_conv_1x1_expand.conv1x1.bias.npy')
groupconv_layer = GroupConvolution(w, b, stride=1, pad=0, groups=groups)
out = groupconv_layer.forward(out)
gamma = np.load(stage +
'0.g_conv_1x1_expand.batch_norm.weight.npy').reshape(
(-1, 1))
beta = np.load(stage + '0.g_conv_1x1_expand.batch_norm.bias.npy').reshape(
(-1, 1))
mean = np.load(stage +
'0.g_conv_1x1_expand.batch_norm.running_mean.npy').reshape(
(-1, 1))
var = np.load(stage +
'0.g_conv_1x1_expand.batch_norm.running_var.npy').reshape(
(-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out_N, out_C, out_H, out_W = out.shape
out = bn_layer.forward(out.reshape(out_C, -1),
train_flg=False).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + 'gconv.txt', out)
out = np.concatenate((residual, out), 1)
#savetxt('./dump/' + 'combine.txt', out)
relu_layer = Relu()
out_N, out_C, out_H, out_W = out.shape
out = relu_layer.forward(out).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + 'stage2.txt', out)
for i in range(1, repeat_num + 1):
residual = out
w = np.load(stage + str(i) + '.g_conv_1x1_compress.conv1x1.weight.npy')
b = np.load(stage + str(i) + '.g_conv_1x1_compress.conv1x1.bias.npy')
groupconv_layer = GroupConvolution(w,
b,
stride=1,
pad=0,
groups=groups)
out = groupconv_layer.forward(out)
out_N, out_C, out_H, out_W = out.shape
gamma = np.load(stage + str(i) +
'.g_conv_1x1_compress.batch_norm.weight.npy').reshape(
(-1, 1))
beta = np.load(stage + str(i) +
'.g_conv_1x1_compress.batch_norm.bias.npy').reshape(
(-1, 1))
mean = np.load(
stage + str(i) +
'.g_conv_1x1_compress.batch_norm.running_mean.npy').reshape(
(-1, 1))
var = np.load(
stage + str(i) +
'.g_conv_1x1_compress.batch_norm.running_var.npy').reshape((-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out = bn_layer.forward(out.reshape(out_C, -1), train_flg=False)
relu_layer = Relu()
out = relu_layer.forward(out).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + str(i) + '_1x1_comp.txt', out)
out = channel_shuffle(out, groups)
#savetxt('./dump/' + 'channel_shuffle.txt', out)
w = np.load(stage + str(i) +
'.depthwise_conv3x3.weight.npy').transpose(1, 0, 2, 3)
b = np.load(stage + str(i) + '.depthwise_conv3x3.bias.npy')
dwconv_layer = DWConvolution(w, b, stride=1, pad=1)
out = dwconv_layer.forward(out)
#savetxt('./dump/' + 'dwconv.txt', out)
gamma = np.load(stage + str(i) +
'.bn_after_depthwise.weight.npy').reshape((-1, 1))
beta = np.load(stage + str(i) +
'.bn_after_depthwise.bias.npy').reshape((-1, 1))
mean = np.load(stage + str(i) +
'.bn_after_depthwise.running_mean.npy').reshape((-1, 1))
var = np.load(stage + str(i) +
'.bn_after_depthwise.running_var.npy').reshape((-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out_N, out_C, out_H, out_W = out.shape
out = bn_layer.forward(out.reshape(out_C, -1),
train_flg=False).reshape(
out_N, out_C, out_H, out_W)
#savetxt('./dump/' + 'after_bn.txt', out)
w = np.load(stage + str(i) + '.g_conv_1x1_expand.conv1x1.weight.npy')
b = np.load(stage + str(i) + '.g_conv_1x1_expand.conv1x1.bias.npy')
groupconv_layer = GroupConvolution(w,
b,
stride=1,
pad=0,
groups=groups)
out = groupconv_layer.forward(out)
gamma = np.load(stage + str(i) +
'.g_conv_1x1_expand.batch_norm.weight.npy').reshape(
(-1, 1))
beta = np.load(stage + str(i) +
'.g_conv_1x1_expand.batch_norm.bias.npy').reshape(
(-1, 1))
mean = np.load(
stage + str(i) +
'.g_conv_1x1_expand.batch_norm.running_mean.npy').reshape((-1, 1))
var = np.load(stage + str(i) +
'.g_conv_1x1_expand.batch_norm.running_var.npy').reshape(
(-1, 1))
bn_layer = BatchNormalization(gamma,
beta,
running_mean=mean,
running_var=var)
out_N, out_C, out_H, out_W = out.shape
out = bn_layer.forward(out.reshape(out_C, -1),
train_flg=False).reshape(
out_N, out_C, out_H, out_W)
#savetxt('./dump/' + 'gconv.txt', out)
out = np.add(residual, out)
#savetxt('./dump/' + str(i) + '_combine.txt', out)
relu_layer = Relu()
out_N, out_C, out_H, out_W = out.shape
out = relu_layer.forward(out).reshape(out_N, out_C, out_H, out_W)
#savetxt('./dump/' + str(i) + '_stage.txt', out)
return out
network = SimpleConvNet(input_dim=(1,28,28),
conv_param = {'filter_num':30, 'filter_size':5, 'pad':0, 'stride':1},
hidden_size=100, output_size=10, weight_init_std=0.01)
# 学習後の重み
network.load_params("params.pkl")
filter_show(network.params['W1'], 16)
img = imread('../dataset/lena_gray.png')
img = img.reshape(1, 1, *img.shape)
fig = plt.figure()
w_idx = 1
for i in range(16):
w = network.params['W1'][i]
b = 0 # network.params['b1'][i]
w = w.reshape(1, *w.shape)
#b = b.reshape(1, *b.shape)
conv_layer = Convolution(w, b)
out = conv_layer.forward(img)
out = out.reshape(out.shape[2], out.shape[3])
ax = fig.add_subplot(4, 4, i+1, xticks=[], yticks=[])
ax.imshow(out, cmap=plt.cm.gray_r, interpolation='nearest')
plt.show()
},
hidden_size=100,
output_size=10,
weight_init_std=0.01)
# 学习后的权重
network.load_params("params.pkl")
filter_show(network.params['W1'], 16)
img = imread('../dataset/lena_gray.png')
img = img.reshape(1, 1, *img.shape)
fig = plt.figure()
w_idx = 1
for i in range(16):
w = network.params['W1'][i]
b = 0 # network.params['b1'][i]
w = w.reshape(1, *w.shape)
#b = b.reshape(1, *b.shape)
conv_layer = Convolution(w, b)
out = conv_layer.forward(img)
out = out.reshape(out.shape[2], out.shape[3])
ax = fig.add_subplot(4, 4, i + 1, xticks=[], yticks=[])
ax.imshow(out, cmap=plt.cm.gray_r, interpolation='nearest')
plt.show()
# pickle 파일에 저장된 파라미터를 cnn의 필드로 로드
cnn.load_params('cnn_params.pkl')
after_filters = cnn.params['W1']
# 학습 끝난 후 갱신(업데이트)된 파라미터를 그래프로 출력
show_filters(after_filters, 16, 4)
# 학습 끝난 후 갱신된 파라미터를 실제 이미지 파일에 적용
lena = plt.imread('lena_gray.png') # imread는 png만 바로 shape으로 적용하게 해준다
# jpg는 PIL를 사용해서 열어야한다
print(lena.shape) #(256, 256) #이미지 파일이 numpy array로 변환됨 -> 2차원
# 이미지 데이터를 convolution layer의 forward()메소드에 전달
# 우리의 convolution 레이어는 오직 4차원만 받음 -> 레나 이미지를 4차원으로 만들어줘야한다
# 2차원 배열을 4차원 배열로 변환
lena = lena.reshape(1, 1, *lena.shape) # *lena.shape = (256, 256)
for i in range(16): # 필터 16개에 대해서 반복
w = cnn.params['W1'][i] # 갱신된 필터w
# b = cnn.params['b1'][i] #갱신된 bias
b = 0 # did not use bias because the image might get distorted
w = w.reshape(1, *w.shape) #3d -> 4d
conv = Convolution(w, b) # convolution 레이어 생성
# 필터 - 학습이 끝난 상태의 필터 (5x5의 작은 필터를 보낸것)
out = conv.forward(lena) #이미지 필터에 적용
# pyplot을 사용하기 위해서 4d-> 2d
out = out.reshape(out.shape[2], out.shape[3])
# subplot에 그림
plt.subplot(4, 4, i + 1, xticks=[], yticks=[])
plt.imshow(out, cmap='gray')
plt.show()
show_filters(before_filters, num_filters=16, ncols=4)
# 학습이 끝난 후 파라미터
cnn.load_params('cnn_params.pkl')
after_filters = cnn.params['W1']
# 학습 끝난 후 갱신된 파라미터 그래프로 출력
show_filters(after_filters, 16, 4)
# 학습 끝난 후 갱신 된 파라미터를 실제 이미지 파일에 적용
lena = plt.imread(
'lena_gray.png'
) # numpy array 로 이미지 열어진다. 단, png 파일만. jpeg 의 경우 외부 패키지 필요
print('lena shape =', lena.shape) # (256, 256) ndarray
# 이미지 데이터를 Convolution 레이어의 forward() 메소드에 전달하기 위해서 2차원 배열을 4차원 배열로 변환
lena = lena.reshape(1, 1, *lena.shape) # * : 튜플 내용을 하나씩 꺼내준다. (256, 256)
print('lena shape =', lena.shape)
for i in range(16): # 필터 16개에 대해 반복
w = cnn.params['W1'][i] # 갱신된 필터
b = 0 # 바이어스 사용 안함
w = w.reshape(1, *w.shape) # 3차원 -> 4차원
conv = Convolution(w, b) # Convolution 레이어 생성
out = conv.forward(lena) # 이미지에 필터 적용
# pyplot 을 사용하기 위해거 4차원을 2차원으로 변환
out = out.reshape(out.shape[2], out.shape[3])
plt.subplot(4, 4, i + 1, xticks=[], yticks=[])
plt.imshow(out, cmap='gray')
plt.show()
